An MCP client is the component inside an AI application that opens and holds one connection to one MCP server. The application you use is the host; it runs one client per server it is connected to. If your app has no MCP client, no MCP server on earth is useful to it.
The distinction matters when something does not work. People say the server is broken when the actual problem is that their application does not support the transport the server uses, or does not support MCP at all.
Host applications with MCP client support include Anthropic's own Claude apps and Claude Code, and coding environments and assistants that have added it. The auth notes recorded on entries in this directory name Claude, ChatGPT and Cursor as the clients whose sign in flows vendors write documentation for, which is a useful signal about where adoption actually is: vendors document against the clients their customers use.
Two people can run the same model against the same server and get different capabilities, because their clients differ in what they expose and how they gate approval. When you are choosing where to build a GTM agent, the client's permission model is worth more attention than the model's benchmark scores. An agent that writes to your CRM with no approval step is a different risk than the same agent asking first, and that choice lives in the client.
This directory records the vendor side of that relationship: which vendors ship a server, what it authenticates with, and whether you can get an account without a sales call. Which client you point at it is your decision and there is no ranking of clients here.
Every number on this page is generated from directory.json at build time and carries the date it was baked: 2026-08-25. Nothing is typed by hand, nothing is rounded, and nothing is estimated. The underlying data is published in full. Where the honest answer is a zero, the zero is printed.